Abstract
|
:
|
This dissertation is an exploratory study of advanced quantitative techniques for assessing culpability for terrorist bombing attacks in Iraq and Israel/Palestine. Using two datasets of terrorist/insurgent activity in Iraq (Jan. 1, 2003--December 31, 2005) and Israel/Palestine (Jan. 1, 2000--December 31, 2005), nine statistical learning classification techniques (simple logistic regression, boosted logistic regression, CART decision tree, J48 decision tree, J48 with AdaBoost.M1, J48 with MultiBoostAB, J48 with Bagging, AdaBoosted J48 with Bagging, and MultiBoosted J48 with Bagging) were modeled through ten scenarios that provide an interpretive context for evaluating the techniques. For each model scenario, the techniques are evaluated for predictive accuracy, true positive prediction rates, kappa statistics, and magnitude of errors. Simple logistic regression served as a baseline model for comparing the statistical learning techniques and paired t-tests corrected for resampling were used to test for statistical significance. All nine statistical and machine learning techniques correctly predicted al-Qaeda-related groups and the specific group al-Qaeda in the Land of the Two Rivers (AQII) with between 70% and 80% accuracy for the various model scenarios. The meta-classifier techniques (those that involved Boosting, Bagging, and combinations of both) significantly outperformed chance and the baseline simple logistic regression model. The statistical and machine learning techniques did not perform well for the Israel/Palestine data. Only models for Hamas predicted significantly better than chance, and these improvements were marginal at best. There were no significant differences between the baseline model and the statistical and machine learning techniques. Policy implications for this study are addressed and discussed. Key words. Terrorism, Statistical learning, Low-intensity conflict, Analysis, Prediction, Culpability, Bombing.
|